NVIDIA B200 GPU Price in India (2026): ₹/hr Cloud Rates, Live

📅 Updated 06 Aug 2026 ⏱ 8 min read 📊 Live pricing feed 💡 Independent comparison
🎛️ B200 Cloud GPU Pricing — Live

NVIDIA B200 GPU Price in India (2026): ₹/hr Cloud Rates, Live

What does an NVIDIA B200 cost per hour in India? As of the latest live snapshot on 06 August 2026, a B200 rents for roughly $4.50 to $5.00 per hour on demand (about ₹428 to ₹475) and about $2.12 to $2.90 per hour on spot (roughly ₹201 to ₹276), across the two live reseller rates gpuprice.in tracks today (Spheron and Nebius). That is the cloud rental per hour, not the hardware price. Buying a B200 outright runs around US$30,000 to US$40,000 per unit, so for most teams renting is the only sensible option.

The short version: India finally has a usable answer, and this page is the live source for it. Prices here are live marketplace snapshots, refreshed daily, and they move. A B200 might cost you $6 one week and $4.20 the next. Below is the full picture, in ₹ and USD, along with where to rent one in India and whether the B200 is worth it over the H100 for your workload.

Before going deeper: this page lists live reseller marketplace rates. It is not a cloud vendor. The numbers come from a real-time feed (parallel.ai), they carry a provider, price type and an as-of date, and every figure is cross-checked against what providers publish on their own price pages. Where those two layers disagree, we show both. See the methodology note at the end.

The live B200 price table: on-demand vs spot

This table is the data layer for the whole article. It is built from the gpuprice.in live snapshot (parallel.ai feed, last updated 06 August 2026, 09:41 UTC, refreshed daily). Prices are live reseller marketplace rates in USD per hour. Where a provider has not listed a B200 rate yet, the cell says "not listed yet."

ProviderB200 on-demand (USD/hr)B200 on-demand (≈₹/hr)B200 spot (USD/hr)B200 spot (≈₹/hr)
Spheron$4.50≈₹428$2.12≈₹202
Nebius$5.00≈₹475$2.90≈₹276
Vast.ainot listed yet~not listed yet~
E2E Networksnot listed yet~not listed yet~
Lambdanot listed yet~not listed yet~
CoreWeavenot listed yet~not listed yet~
RunPodnot listed yet~not listed yet~
AWS / Azure / GCPnot listed yet~not listed yet~

*Source: gpuprice.in live snapshot via parallel.ai, last updated 2026-08-06 09:41 UTC, refreshed daily. Prices are live reseller marketplace rates in USD/hr. Rupee figures are conversions at ~₹95/USD, not provider-published INR list prices.*

For context, the same live snapshot tracks the older workhorses: H100 on-demand from $2.50/hr and A100 80GB on-demand from $1.67/hr on the cheaper marketplace providers. We keep this article B200-focused, so those are side notes. See our full H100 vs H200 vs B200 comparison for the chip-by-chip pricing picture.

Live marketplace vs provider list prices

Here is an honest caveat you will not get from most pricing aggregators. The gpuprice.in snapshot can sit below what providers publish on their own pages. Two real examples as of 06 August 2026:

Our snapshot and their pages move independently and can disagree by as much as a few dollars. We show both layers so you are never misled into thinking the lowest number is guaranteed. Treat the table as the live marketplace check, and the provider page as the anchor for a firm commitment.

How much does an NVIDIA B200 GPU cost per hour to rent?

On the live marketplace today, you can rent a B200 for $4.50 to $5.00/hr on demand (about ₹428 to ₹475) and $2.12 to $2.90/hr on spot (about ₹201 to ₹276). Those are the two tracked reseller rates, both as of 06 August 2026. The on-demand figure is the one to plan around; spot is a discount you should only assume when you can tolerate interruptions.

A couple of layers matter here. Live marketplace rates (what gpuprice.in tracks) can sit below a provider's own published list price, and both move daily. Nebius, for example, lists its B200 at $7.15/hr on demand on its own page, while the marketplace snapshot has it at $5.00/hr. If you see different numbers elsewhere, compare the date and the layer, not just the number.

To put it in Indian terms: at ~₹95/USD and roughly 730 hours a month, a B200 running on demand for the full month costs about ₹3.1 lakh to ₹3.5 lakh. On spot, the same month is around ₹1.5 lakh to ₹2 lakh, if you can live with preemption. Most teams should compare this against the new-build cost of ~$30,000 to $40,000 per card before committing to anything.

B200 vs H100: what's the price difference per hour?

The B200 costs more per hour and delivers much more per hour, so the simple price gap understates the real value gap. On the live snapshot, B200 on-demand sits at $4.50 to $5.00/hr, while H100 on-demand starts near $2.50/hr (Vast and Spheron). In ₹: about ₹475 vs ₹238 per hour at on-demand rates.

But the B200 is not "just" 2x the price of the H100. NVIDIA positions it as roughly 15x faster than the H100 for inference and 3x for training on the DGX platform, and third-party figures put the per-something math even further in its favour. In practice, that means an Indian team doing serious inference or frontier fine-tuning often gets more compute per rupee on a B200 even though the hourly bill is higher. For simple fine-tuning or budget workloads you get more value from the cheaper card. Our H100 vs H200 vs B200 comparison lays out the full three-way table including price, VRAM, memory bandwidth and FP4 throughput.

B200 vs H200: what's the price difference per hour?

The H200 is the cheaper Blackwell-adjacent middle option, not the fastest, and the gap to the B200 is smaller than the gap to the H100. The live snapshot has H200 on-demand at $2.80 to $4.31/hr across Spheron, Vast and RunPod (some providers list it, some do not), versus $4.50 to $5.00/hr for the B200. In rough terms you pay a 15 to 70 percent premium for the B200 over the H200 depending on which provider you price against.

Here is why the premium exists. The H200 carries more HBM3e memory than the H100 but does not reach the B200's FP4 throughput. If your training or inference is memory-bound rather than raw-FLOPs-bound, the H200 is often the smarter buy for Indian budgets. If you need maximum tokens per dollar on large inference workloads, the B200 wins. We break down the trade-off in detail in the shorter H200 vs B200 sections above and the full comparison in our H200-vs-B200 pricing analysis.

Where can I rent a B200 GPU in India?

For an Indian customer, the realistic options divide into global marketplace providers you reach from Mumbai, and domestic providers with in-country capacity. As of 06 August 2026, here is the practical list:

Latency matters and changes the calculus. Most cheap B200 marketplace capacity still routes through the US or EU or Singapore, and a Singapore-region instance is about 30 to 50 ms from Mumbai. That is fine for training and most batch inference, but it is a real consideration for latency-sensitive real-time workloads. For domestic compliance and residency needs, E2E and Cyfuture become the more interesting options. If you are already choosing between Indian providers for H100 capacity, our E2E Networks and Vast.ai India guide explains how the Indian and global layers work together.

Is B200 available in Indian data centers?

Broadly yes, with the caveat that it is still ramping. As of mid-2026:

The honest caveat is that a lot of the cheap B200 capacity you will find on the open market still sits abroad (Singapore, EU, US). If your requirement is regulatory compliance, data residency or simply a latency-friendly connection, the domestic options above are the ones to shortlist. If your requirement is lowest cost per hour and latency is not critical, the global marketplace rate wins. We update the domestic availability list as providers surface, so check back periodically rather than treating this as a static answer.

What's the B200 GPU price in India per hour (in ₹)?

The B200 rents for about ₹428 to ₹475 per hour on demand and roughly ₹201 to ₹276 per hour on spot, at the current ~₹95/USD conversion. These are conversions from live USD marketplace rates, not published INR list prices. No provider publishes a static per-hour INR rate for the cloud B200 as of the update date, and we flag that difference explicitly.

The one firm rupee figure with government weight is the IndiaAI tender. In the BharatGen / IndiaAI Mission fourth-round tender, the hourly B200 rate came down about 10% to ₹290.70 for a single unit (and ₹2,325.60 for an 8-unit setup), down from ₹323 and ₹2,584 respectively. That is a tender reference price, not an open-market rate, but it is a durable, citable Indian rupee anchor. For comparison, B300 rates in the same tender ran ₹351 to ₹2,808/hr depending on setup.

One important distinction: the subsidized IndiaAI compute often quoted at ₹65 to ₹92/hr is H100 and H200 class, not B200. Do not confuse those numbers with B200 pricing. The B200 tender rate is the ~₹290.70 figure above.

Power is the second rupee note. A B200 runs at roughly 1000W (SXM), and Indian industrial electricity tariffs run about ₹4.50 to ₹7.20 per kWh. So the electricity alone to run a single B200 card for an hour is roughly ₹4.50 to ₹7.20 just for the card, before cooling, network and overhead. A full DGX B200 system draws ~14.3kW, which is a materially larger power bill. Work those into your total cost of ownership, not just the hourly rent.

What's the cheapest way to rent a B200 GPU in 2026?

The cheapest tracked B200 rental in 2026 is Spheron spot at $2.12/hr (about ₹202/hr) in the live snapshot. If you compare across the market:

Two important things to know before you chase the cheapest number. First, spot means preemptible: the provider can reclaim the machine for another customer, so you must design for interruptions (checkpoints, restartable jobs). Second, live marketplace spot rates are volatile and can sit well below provider list prices, as discussed in the table section, so confirm the live number at the moment you commit. If you need guaranteed availability for a production workload, take on-demand and accept the premium. For experiment, fine-tuning and off-peak batch work, spot is where the savings live.

Is B200 worth it over H100 for an Indian startup?

For most frontier inference and serious training, yes; for budget fine-tuning, not necessarily. The decision turns on throughput per rupee rather than the headline hourly number.

Independent analyses of the B200 versus the H200 and H100 are striking on throughput. SemiAnalysis' InferenceX work found the B200's NVFP4 path can deliver up to about 2.95x the performance per dollar versus the H200 INT4 on workloads like Kimi, and about 2.38x cheaper per token than the H100 on MiniMax M2.7-style inference. In plain terms: on token-heavy inference, the B200 can cut your effective cost even though its hourly bill is higher.

That does not make it right for everyone. Realistic guidance for an Indian team:

A single 7B or even 70B fine-tune does not need a B200. A monthly bill of ₹8 to ₹10 lakh is hard to justify on the hobby tier. We cover how to pick between chip classes in our H100 vs H200 vs B200 comparison.

B200 spot price vs on-demand: which should I use?

On demand, you pay for guaranteed availability; on spot you save roughly half but risk preemption. On the live snapshot, the B200 spot price is $2.12/hr on Spheron and $2.90/hr on Nebius, versus on-demand $4.50/$5.00/hr respectively. That is a 45 to 53 percent saving on the marketplace rate.

Use on-demand when the job cannot be interrupted: a production inference endpoint, a client deadline, a long training run you cannot afford to lose. Use spot when you can handle the provider reclaiming the machine: training with regular checkpoints, off-peak inference, hyperparameter sweeps, experimental runs.

For a ~1000W GPU in India there is a power angle too. A single B200 card draws ~1000W; at Indian industrial electricity tariffs of roughly ₹4.50 to ₹7.20 per kWh, the idling-vs-computing hours and how long you keep a spot box warm both hit your bill. Spot hour pricing already prices in that the machine may return to the idle queue, which is part of why it is cheaper. For a fuller treatment of how on-demand and spot differ on the H100, where the patterns are identical, see our spot vs on-demand pricing guide.

B200 vs H200 vs H100: which should I rent in India?

A quick decision guide, price and role based:

NeedChooseLive snapshot rate (06 Aug 2026)
Max inference throughput, token-heavyB200$4.50–5.00 on-demand; $2.12–2.90 spot
Memory-bound training, lower premiumH200$2.80–4.31 on-demand
Budget fine-tuning, small teamsH100from $2.50 on-demand; spot from $0.99
First experiments, subsidised accessIndiaAI H100/H200 (not B200)₹65–92/hr

The B200 is the fastest and most expensive. The H200 is the memory-heavy middle option. The H100 is the budget workhorse. If your job is token-heavy inference at scale, rent a B200. If it is memory-bound training and you want to save, the H200 is the smart choice. If you are fine-tuning a small model or just starting, rent an H100 on spot first.

Your best first move in most cases is to start on spot H100 (or IndiaAI subsidised H100/H200) to validate the workload, then move to the B200 once throughput actually becomes the bottleneck. Nothing about hardware spending rewards guessing early. Our H100 price trend through 2026 shows how the lower tier has moved, and our DGX station vs renting cloud GPUs guide helps you decide between owning boxes and renting cloud time, which matters for an Indian team given power and import considerations.

How much does it cost to fine-tune a model on a B200 (student / small team)?

A single small fine-tune does not cost much, and it almost never needs a B200. Here is a quick worked budget for a data point:

JobHoursB200 on-demand (≈$4.75/h avg)H100 on-demand (≈$2.50/h)H100 spot (≈$0.99/h)
LoRA on 7B1–2₹452–904₹238–475₹94–188
Fine-tune 7B4–8₹1,808–3,615₹951–1,902₹376–753
Fine-tune 70B20–50₹9,036–22,592₹4,753–11,882₹1,880–4,700

*(Rupee figures are conversions at ~₹95/USD. Actual hours vary widely with dataset size, batch size and framework.)*

A student or small team fine-tuning a 7B or 70B model is usually better served by a spot H100 (about ₹376 to ₹4,700 total for a 70B run) than a B200. You pay more per hour on the B200 and rarely need the extra throughput at this scale. Reserve the B200 for throughput-bound inference or very large training runs. If you are cost-constrained, the IndiaAI subsidised H100/H200 compute at ₹65 to ₹92/hr is the cheapest legitimate path and beats any open-market card for budget work.

Which B200 cloud provider has the best latency from India?

For an Indian-based user, latency depends on region, and the fastest option is whichever is physically closest to you. In practice:

For batch training, inference pre-processing and data pipelines, the 30 to 50 ms latency of a Singapore box is irrelevant. For real-time user-facing inference, every millisecond counts and a domestic data centre wins. The best B200 provider from India on pure latency is a domestic one (E2E or Cyfuture); the best on pure price is a global marketplace provider (Spheron spot). Pick based on which constraint the workload cares about, not on a single number.

India angle, import and hardware cost context

Beyond hourly cloud rates, three India-specific facts shape how the B200 plays out here.

Hardware cost. A B200 unit was guided at roughly US$30,000 to US$40,000 at launch (Jensen Huang, GTC 2024). That guidance is dated, but it anchors the purchase-vs-rent decision: buying one B200 card and stuffing it into a machine you build is a multi-crore-rupee-plus undertaking once you add the server, networking, power and cooling. Renting at ₹4 to ₹5 lakh a month is the financially sensible default for nearly everyone. Comparing the two paths is exactly the question our DGX station vs renting cloud GPUs guide answers.

Power. At roughly 1000W per SXM B200, and ~14.3kW for a full DGX B200 system, power is a first-class cost in India where industrial electricity runs ₹4.50 to ₹7.20 per kWh. For a domestic operator, that is real money on a 24/7 box and one reason domestic providers lease compute rather than sell boxes.

Import and compliance. AI servers sold and used in India require BIS registration, and Indian data centre operators are investing heavily (E2E's Chennai cluster, Yotta's Blackwell Ultra buildout) as part of the sovereign-AI push. We do not quote a specific import-duty percentage here because no verified figure is available; that is an honest gap. If you are importing a B200-class server yourself, confirm the current duty and BIS position with a customs or regulatory professional rather than relying on this page. The market reality is that India's AI infrastructure is scaling fast, and that is a large part of why demand for B200 class compute keeps climbing; see our note on why global GPU demand keeps climbing for the demand-side context.

How we collect these prices (methodology)

Every price on this page comes from the gpuprice.in live pricing feed, which pulls reseller marketplace rates through the parallel.ai API. The feed refreshes daily (last update 06 August 2026, 09:41 UTC) and records on-demand and spot rates per provider per GPU. Where a provider has not surfaced a rate yet, the cell is marked "not listed yet" rather than filled with a guess.

Two things to know about the feed. First, it tracks live reseller marketplace snapshots, which is why they can sit below a provider's own published list price; we show both layers so you always see the real range. Second, rupee values on this page are conversions at ~₹95/USD, not provider-published INR list prices, and we flag that everywhere. Every price in this article carries a provider, a price type (on-demand or spot) and an as-of date of 06 August 2026 unless a different date is stated.

Monetisation disclosure

gpuprice.in is an independent pricing and comparison site. It may earn a fee or referral from some of the providers it lists if you sign up or rent through the links on this page. This does not change the price you pay, and it never changes which providers we cover or how we rank them. Listed figures are live marketplace snapshots and can differ from a provider's own page; always confirm the final rate on the provider's site before committing. We are transparent that this is a listing and aggregation service, not a cloud vendor, and that our rankings are data-led, not sales-driven.

Frequently asked questions

Short answers below; the live table near the top has the current numbers, and each topic above has the fuller reasoning.

How much does an NVIDIA B200 GPU cost per hour to rent?

Roughly $4.50 to $5.00/hr on demand and $2.12 to $2.90/hr on spot as of the live table (06 Aug 2026). See the live table above for current figures.

B200 vs H100: what's the price difference per hour?

The B200 runs about $2 more per hour on demand than an H100, but delivers far more throughput, so the value gap is smaller than the price gap. Full table in the B200 vs H100 section.

B200 vs H200: what's the price difference per hour?

The H200 is typically 15 to 70 percent cheaper than the B200 on demand, and is the smarter buy for memory-bound training. See the B200 vs H200 section.

Where can I rent a B200 GPU in India?

Spheron and Nebius on the global marketplace, AWS P6-B200 for enterprise, and E2E Networks and Cyfuture AI for in-country capacity. See the provider section above.

Is B200 available in Indian data centers?

Yes, it is ramping: E2E Networks procured 1,024 B200s for Chennai, Cyfuture rents B200 servers from an Indian DC, and Yotta is building a large Blackwell Ultra fleet.

What's the B200 GPU price in India per hour (in ₹)?

About ₹428 to ₹475/hr on demand and ₹201 to ₹276/hr on spot at ~₹95/USD; the IndiaAI tender reference for a single B200 is ₹290.70/hr. See the INR section.

What's the cheapest way to rent a B200 GPU in 2026?

Spheron spot at $2.12/hr is the lowest live-tracked rate. Spot is preemptible, so design around interruptions. See the cheapest-B200 section.

Is B200 worth it over H100 for an Indian startup?

For token-heavy inference and frontier training, yes; for budget fine-tuning of 7B or 70B models, an H100 or H200 on spot gives more value per rupee.

B200 spot price vs on-demand: which should I use?

Use spot for checkpointed training, off-peak inference and experiments; use on-demand for production inference and jobs that cannot be interrupted.

B200 vs H200 vs H100: which should I rent in India?

Pick the B200 for max inference throughput, the H200 for memory-bound training, and the H100 for budget fine-tuning. Start on spot H100 and scale up only when throughput demands it.

Verified sources

Primary, first-party, URL-verified on 06 August 2026:

Secondary, high value:

Performance-per-dollar figures referenced from the aisearch handoff (SemiAnalysis InferenceX) for the B200-vs-H100/H200 analysis; verify the live URL before publishing that specific link.

*Last updated: 06 August 2026. Prices are live marketplace snapshots and change daily. For firm commitments, always confirm on the provider's own page.*